Deep Learning Based Automated Chest X-ray Abnormalities Detection

Abnormalities related to the chest are a fairly common occurrence in infants as well as adults. The process of identifying these abnormalities is relatively easy but the task of actually classifying them into specific labels pertaining to specific diseases is a much harder endeavour. COVID-19 sufferers are multiplying at an exponential rate, putting pressure on healthcare systems all around the world. Because of the limited number of testing kits available, it is impractical to test every patient with a respiratory ailment using traditional methods. Thus in such dire circumstances, we propose the use of modern deep learning techniques to help in the detection and classification of a number of different thoracic abnormalities from a chest radiograph. The goal is to be able to automatically identify and localize multiple points of interest in a provided chest X-ray and act as a second level of certainty after the radiologists. On our publically available chest radiograph dataset, our methods resulted in a mean average precision of 0.246 for the detection of 14 different thoracic abnormalities.